Maya 2 TTS 評測：涵蓋十一種印度語言，資料夾名稱仍是 maya3

The model is called Maya 2. The folder is called maya3.

That is not a typo i made. The Hugging Face card says it plainly: "the existing package, Python class, and configuration filename remain maya3, Maya3TTS, and maya3_config.json."

So you download a model named after one version, and import a class named after another. There is a version three somewhere in this company's history that either got renamed or got skipped, and nobody cleaned it up before the launch.

i found that oddly reassuring. A seed-stage Indian startup shipping an Apache 2.0 voice model does not have time to rename a config file. It has time to ship.

Maya Research released the open weights for Maya 2 Native on 25 October 2026. Twelve languages, eleven of them Indian.

Apache 2.0, which means commercial use, fine-tuning, the whole thing.

Dheemanth Reddy, the CEO, posted on X that Maya 2 beats ElevenLabs v3, Cartesia Sonic 3.5 and Bulbul v3 across Indian languages.

i wanted that to be true. Then i looked at where the benchmark came from.

What actually shipped

The model card is unusually specific, which i respect.

The language list covers Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and English. That is the hardest part of this problem and they went straight at it.

Cross-lingual cloning is the feature i did not expect. A Hindi reference clip can produce Tamil speech in the same voice.

Not a translation demo. Same speaker, different language.

If that holds up in practice, it is the most useful thing in the release.

The benchmark problem

here's a question people always ask when a small lab claims a win: who ran the test?

On VoiceArena's Hindi leaderboard, Maya 2 Native sits at rank two with an Elo of 1047. Cartesia's Sonic 3.6 is above it at 1094. Sarvam's Bulbul v3 is well below at 957.

Those are third-party numbers with sample counts and confidence intervals attached.

Then there is the claim that Maya 2 beats ElevenLabs v3. That comparison came from the CEO's own post, and the post does not show how the test was run.

i do not think that is dishonest. Launch posts do this constantly. But it means one number in the stack is audited and the other is a press release.

And on Artificial Analysis, which tracks 103 speech models globally, Maya 2 Global sits at 1048 Elo. That is rank 28. ElevenLabs v3 is above it, and so is most of the field.

So the honest version is narrower than the tweet. Maya 2 is genuinely competitive inside Indian languages. It is not competitive on a global leaderboard, and it was never going to be at this size.

The latency numbers are the real story

the first time i tried building a voice agent for Hindi, the problem was not quality. It was the pause.

Every hosted model added a beat between the user finishing and the agent starting. In English you barely notice. In a loan-repayment call in Marathi, half a second of silence reads as a dropped connection or a scam.

Maya claims 75 ms time to first byte, with 7.3 ms of variance between invocations, measured from Bangalore to their Mumbai region. That is a real number for a real use case.

The docs have a better detail buried in them:

First audio, new connection per request: 389 ms
First audio, connection reused: 88 ms

Four times slower if you open a fresh connection every time. Almost nobody reads that line before shipping, and then they wonder why their demo feels sluggish.

The API returns raw PCM, 16-bit little endian, mono, 24 kHz, with no file header. Write those bytes straight to a .wav and no player will open it. Their own docs warn about this in a way that suggests someone did it, filed a bug, and got told it was their fault.

The small things that will bite you

what actually happens is that voice names are case-sensitive.

The docs show voice: "Ananya" and voice: "Arjun". Lowercase it and you get different audio, because the model interpolates whatever string you hand it. There is an error example in the docs with "ananaya" spelling.

That is not a bug. It is a design choice that will cost someone an afternoon.

Emotion tags exist but live in a separate model. You have to set model: "Maya 2 Native Emotional" and write [excited] or [whispers] inline.

Ten tags total. The default model ignores them.

And code-mixed text, the thing Indian users actually type, is handled by leaving the language field out and hoping. The docs say auto-detect handles it. i would want to test that on a sentence that switches twice.

A detour about keyboards

India has about 1.5 billion people and 22 official languages, and typing most of them on a phone is miserable.

i have watched relatives give up on a Gujarati message and switch to English because the keyboard was fighting them. Not because they cannot read Gujarati. Because thumb-typing it is slow enough that the conversation moves on without them.

That is the actual market argument for voice, and it is stronger than any Elo table.

It is also why this release matters more than a normal TTS drop. Voice is not a nice accessibility feature in that context. It is the only input method that does not punish you for your first language.

Who should actually download this

Here is the part nobody wants to say out loud.

Five gigabytes of BF16 weights is not open in any practical sense for the developer this model is aimed at.

An Apache 2.0 licence and a 5 GB checkpoint mean nothing if you are on a laptop with 8 GB of RAM and no GPU. You can download it. You cannot run it.

The real access path is still the API. Open weights here are a promise to startups and research labs with hardware, not to a student in Pune with a MacBook Air.

If you are building a quick prototype, use the hosted endpoint and skip the weights entirely. If you are fine-tuning for one dialect, the weights are exactly what you need and the licence is genuinely generous. If you are building a commercial product and your users are on cheap Android phones, neither path solves your problem today.

That last group is most of the market they are targeting. That is the gap.

The bit i keep coming back to

The maya3 folder name is still the detail i like most.

Big labs ship clean. Everything named, everything consistent, a brand guide for the config files. Small labs ship the thing that works and file the rename for later.

Maya took $1.9M in seed money and put out an Apache 2.0 model covering eleven Indian languages, which is more than most funded Western labs have bothered to do in any language that is not English.

The benchmark claims need a third party to repeat them. i would rather see that than another launch post with a chart.

Meanwhile there is a file on Hugging Face that says Maya 2 on the outside and maya3 on the inside. Somebody should fix it. i also hope nobody does for a while.